The problem
The knowledge is usually already written down. What fails is reaching it.
A clinical officer at 2am has the same WHO guidance available as anyone else — as a PDF or a web page nobody can search under pressure. So people ask a chatbot instead, and a chatbot's failure mode is the dangerous one: it is fluent, it is confident, and it will produce a paediatric dose that appears in no document anywhere.
The fix is not a better-sounding model. It is a tool that can only repeat what a named source actually says, and that visibly refuses when the source is silent.
What it does
- You ask a question.
- It retrieves candidate passages from a store of 292 passages drawn from 22 WHO fact sheets.
- The agent must quote the matching passage verbatim and name its source and section.
- If nothing stored answers the question, it returns
NOT COVERED - this guideline does not address that, and the interface shows that as a first-class result, styled with the same weight as an answer — not as an error.
Every question, answer, quoted passage and refusal flag is written to a log table, so the refusals are as inspectable as the answers.
Why it is built in Momen
The whole application — relational tables, the AI agent, the Actionflow that runs it, and four pages — is assembled in Momen's visual builder. No source code exists for this project; there is no repository to link because there is nothing to check out.
That is the honest claim for the no-code prize: not "AI helped me write code", but an app whose data model, agent configuration, logic flow and interface were all produced inside Momen.
The corpus was prepared outside Momen — WHO pages fetched, stripped, and chunked at sentence boundaries into a CSV — and loaded through Momen's own CSV import.
On the free tier, stated plainly
This was built on Momen's Free plan, and two limits shaped what shipped. Both are on the plan page; neither is a bug.
Vector storage requires the Basic plan. The design called for semantic retrieval — embeddings on the passage text, cosine distance against the question. The platform gates that behind a paid tier, so retrieval in the shipped app is lexical rather than semantic. A question phrased in the guideline's own words retrieves well; a question phrased differently retrieves worse. This is the single biggest limitation of the submitted version and the first thing that would change.
Publishing to a public URL requires the Basic plan. The free tier is preview-only, so the demonstration is a recording of the working app rather than a link you can open.
Momen's Copilot also reported it had enabled vector search when the platform had in fact refused the request. Worth knowing if you build on it: verify what the builder tells you it did.
What it does not do
- It does not diagnose. It retrieves published text.
- It does not paraphrase a clinical fact, and it will not give a dose that is not word for word in a stored passage.
- WHO fact sheets are public-health summaries, not clinical guidelines — they carry no dosing, so most dosing questions are correctly refused rather than answered.
- Retrieval accuracy is unmeasured. There is no held-out question set and no figure for how often the right passage comes back. Saying otherwise would be the easy lie in a project whose entire argument is about not making things up.
- It has never been used by a clinician.
What I learned
That the interesting engineering in a retrieval tool is the refusal, and that a no-code builder can express it perfectly well — the constraint lives in the agent's system prompt and in how the interface treats "no answer" as a result rather than a failure.
Also that a builder's own AI will tell you it did something it did not do. The claim that vector search was enabled looked completely plausible in the transcript and was false on the platform.
What's next
Semantic retrieval on the Basic tier, which is the one change that would most improve it. A measured retrieval evaluation — a set of questions with known correct passages, and a published hit rate. And swapping the fact-sheet corpus for a real clinical guideline, at which point dosing questions become answerable from the document instead of refused.
Log in or sign up for Devpost to join the conversation.